Claude Opus 5

Claude Opus 5 launched 24 July 2026 with a 1M-token context, 128k output and $5/$25 per million tokens. What it is best at, and where it is overkill.

Claude Opus 5 is Anthropic's general flagship model, released 24 July 2026. It costs $5 per million input tokens and $25 per million output tokens, takes a 1,000,000-token context, returns up to 128,000 output tokens, and has a May 2026 knowledge cutoff — the most recent of any Claude model. Anthropic recommends it as the default starting point for complex agentic coding and enterprise work.

It is not the top of the range: Claude Fable 5 sits above it at twice the price. But Opus 5 is the newest brain in the lineup, and for work that needs more than Sonnet 5 it is the model to reach for. This page covers what it costs, how the effort parameter behaves, the API changes that break older code, and where paying for it is a waste.

#Claude Opus 5 pricing in full

Opus 5 costs $5/$25 per million tokens on the standard API. That is the same headline price as the whole Opus 4.x line, and a third of what the retired Opus 4.1 charged at $15/$75. Every discount stacks on the base rate rather than replacing it.

If you are sizing a budget rather than a single call, the pricing page covers subscriptions and API rates side by side, and the Claude API page works caching and batching through with worked numbers.

#How the effort parameter works on Opus 5

Thinking is on by default on Opus 5. You do not enable it; you tune how much of it happens, using an effort ladder with five rungs. The wire shape is output_config, not the old thinking.budget_tokens field:

{
  "model": "claude-opus-5",
  "max_tokens": 8192,
  "output_config": { "effort": "max" },
  "messages": [{ "role": "user", "content": "Refactor this module." }]
}

The distinction Anthropic draws is worth internalising: the thinking parameter controls whether Claude produces thinking blocks before answering, while effort controls how much work goes into the whole response, thinking included. They are not the same dial. Thinking tokens are billed as output tokens, count against max_tokens, and count toward your rate limits even when the display is omitted. The mechanics across the whole lineup are on the thinking modes page.

#API parameters that no longer work

Three sampling parameters are dead on Opus 5. temperature, top_p and top_k have been deprecated since Opus 4.7, and passing any of them a non-default value returns a 400 error rather than being silently ignored. Anthropic's replacement advice is blunt: omit them and steer behaviour through the prompt instead. Assistant message prefilling — starting the model's reply for it — was removed from 4.6 onward and also 400s.

That removes most of the classic determinism levers. If you were setting temperature: 0 for reproducible evaluation runs, there is no equivalent knob; you get variance and must design evaluations that tolerate it. That is a real regression for anyone running strict suites. Practical patterns are covered on the Claude for developers page.

Two additions run the other way. On the Batch API, the output-300k-2026-03-24 beta header raises max_tokens to 300,000 on Opus 5 — as well as Opus 4.8, 4.7, 4.6 and Sonnet 5 — at standard batch pricing. Anthropic notes a single 300k-token generation can take over an hour, so pair it with the 1-hour cache TTL. And a separate beta, mid-conversation-tool-changes-2026-07-01, lets you add or remove tools between turns on Opus 5 without invalidating the prompt cache, which matters for MCP-driven agents whose toolset changes as a task progresses.

#Migrating from Opus 4.x to Opus 5

For most codebases this is a one-line change:

- model="claude-opus-4-8"
+ model="claude-opus-5"

Then check three things.

1. Thinking is on by default now. Opus 4.8 has adaptive thinking available but requires configuration; Opus 5 has it on out of the box. Your token accounting will shift upward if you were not budgeting for thinking output.
2. The xhigh/max plus disabled-thinking combination 400s, as above. Audit any config that sets both.
3. Delete carried-over verification instructions. Anthropic explicitly calls out prompt lines such as "include a final verification step" or "use a subagent to verify" — habits from earlier Opus versions — as causing over-verification on Opus 5. The model does this work itself now; instructing it again makes runs longer and more expensive for no gain.

Coming from an older generation is a bigger job. Anything written for Claude 4 or 4.5 also carries temperature, prefills and manual thinking budgets, all of which now fail. Claude Code ships a /claude-api migrate skill that applies model-ID swaps and breaking parameter changes across a repository. Historical context on the earlier releases is in the Opus family archive.

#Rate limits for Claude Opus 5

Anthropic's tiers are named Evaluation, Start, Build, Scale and Custom — not Tier 1 to 4 — and organisations are placed automatically based on usage history and account standing. Opus 5 gets its own rate-limit bucket, separate from the shared Opus 4.x pool, so migrating to it does not eat the quota your 4.x traffic is still using.

Two behaviours are worth designing around. Cache reads do not count toward your input-tokens-per-minute limit — only fresh input and cache writes do — so aggressive caching buys throughput as well as money. And hitting a spend cap returns a 429 with error_code: "enforced_spend_limit_reached" and no retry-after header, which will confuse a naive backoff loop into retrying forever. A self-imposed limit returns a 400 instead. Fable 5's limits are much tighter than these at every tier.

#How capable is Opus 5, honestly?

Independent measurement puts it at the top. On the Artificial Analysis Intelligence Index, Opus 5 ranks first with a score of 63. On ARC-AGI-3, verified by the ARC Prize foundation, it scores 30.16% against GPT-5.6 Sol's 7.78% — a roughly four-fold gap on novel-problem reasoning.

Anthropic's own claims are relative rather than numeric. On Frontier-Bench v0.1 it says Opus 5 "surpasses all other models" and more than doubles Opus 4.8's performance at lower cost per task; on CursorBench 3.2 at max effort it lands "within 0.5% of Fable 5's peak score, but at half the cost per task"; on Zapier's AutomationBench it claims a pass rate around 1.5× the next-best model at equal cost.

Where Opus 5 loses is also documented. It generates no images, video or audio and accepts neither video nor audio input. Claude's long-context retrieval is the weakest of the big three despite the 1M window. And rival models hold the top scores on OSWorld 2.0 and Terminal-Bench. The full head-to-heads are on Claude vs ChatGPT and Claude vs Gemini.

#When Claude Opus 5 is overkill

Opus 5 is five times the input price of Sonnet 5 and 2.5× its output price. That premium is worth paying for a narrower set of tasks than most teams assume. Skip it when:

The honest default remains Sonnet 5, with Opus 5 reserved for complex agentic coding, large refactors, vision-heavy work, computer use, and anything where knowledge past January 2026 matters. If you want that framed as cost per completed task rather than cost per token, the model comparison does the arithmetic, and the model index lists every current and retired option. Teams with procurement constraints should also check what is bundled per seat on the enterprise page before buying API credit.

#Frequently asked questions